A major artificial intelligence conference rejected 497 papers in early 2026. Not because the research was bad. Because the people reviewing the research used AI to write their reviews and got caught.
The conference organizers had hidden watermarks in the papers distributed to reviewers. When they analyzed the submitted reviews, they found the watermarked language coming back in the feedback. The reviewers hadn’t read the papers. They’d fed them to an AI and submitted whatever came out.
497 papers. Roughly 2% of all submissions. Caught in a single sweep.
How many weren’t caught?
What Peer Review Is Actually For
Peer review is the process by which scientists check each other’s work before it gets published. It’s the thing that separates science from opinion. A researcher submits a paper, other experts in the field evaluate the methodology, challenge the conclusions, and decide whether the work is solid enough to add to the scientific record.
It’s imperfect. It’s slow. It’s sometimes biased. But it’s the mechanism that keeps outright nonsense out of journals.
Now it’s being defeated by AI writing reviews that sound like expert analysis but are generated in seconds by a system that hasn’t actually evaluated anything. The reviewer gets paid in academic credit. The paper gets a review. The journal publishes it. The scientific record grows. And somewhere in that record is research that was never actually checked by a human who understood it.
It’s Not Just the Reviews
The peer review fraud is visible because of the watermark trick. The harder problem is the research itself.
Scientists are publishing AI-generated papers. Not using AI as a writing assistant. Using AI to produce results, generate figures, and write conclusions for experiments that may have been conducted minimally or not at all. The AI output looks like research. It has the structure, the citations, the statistical language. It passes automated plagiarism checks because it’s not copied, it’s generated.
Some journals have started requiring authors to declare AI use. Most researchers in surveys say they use AI in their work. Almost none specify what “use” means. There is no standard. There is no enforcement. There is a growing body of published literature where nobody is certain how much of it reflects actual experiments.
Why This Matters Beyond Science
Medical treatments get approved based on research. Engineering standards get set based on research. Government policy gets shaped by research. The drugs your doctor prescribes, the materials in the buildings you work in, the safety regulations on the products you use — all of it traces back to a research literature that is now being infiltrated by AI-generated content that may or may not correspond to anything real.
This isn’t abstract. In 2023, a paper citing fake legal cases generated by ChatGPT was submitted in a real court. The citations looked real. The cases didn’t exist. The lawyer who submitted it didn’t check. The judge caught it.
In science, there’s no judge. The watermark trick works once, in one conference, for reviewers who left an obvious trace. For everything else, the detection systems are nowhere close to keeping up with the generation systems.
The System Assumed Human Incentives
Academic integrity systems were built on the assumption that fraud required effort. Fabricating data is hard. Writing fake reviews takes time. The risk of getting caught was high enough that most people didn’t bother.
AI changed the effort equation. Generating a convincing fake review takes ten seconds. Writing a paper that looks real takes an afternoon. The risk hasn’t changed. The cost has dropped to almost zero.
When the cost of fraud collapses, you need either much better detection or much higher consequences. Neither is in place. And the scientific record, which took centuries to build, is being quietly corrupted in real time.
The 497 caught papers are not the story. They’re the ones we know about.